1. Encoding of Stimulus Probability in Macaque Inferior Temporal Cortex
Andrew H. Bell, Christopher Summerfield, Elyse L. Morin, Nicholas J. Malecek, Leslie G. Ungerleider 
Current Biology 
Volume 26, Issue 17, Pages (September 2016)
DOI: /j.cub
Copyright © Terms and Conditions

2. Figure 1 Behavioral Performance in Delayed Match-to-Sample Task
(A) Monkeys were trained on a delayed match-to-sample task that required them to identify which of two stimuli (a face or fruit) best matched a previously presented cue. Cue stimuli were pseudorandomly selected on a session-by-session basis from a set of eight possible images for each category and were degraded at two different levels by the addition of Gaussian noise to the image (low and high noise trials) (N.B. this panel shows an example of a low-noise cue). The trials were arranged into five blocks, (presented in random order), each with a different probability of the cue being a “face” versus a “fruit” (0%, 25%, 50%, 75%, 100%).
(B) Left: a robust behavioral effect emerged whereby the subject’s choice was strongly biased by prior probability—“face” was chosen more often than “fruit” when trial proportions were biased toward faces and vice-versa. This trend was found at both noise levels. Right: hit rates and false alarms were more influenced by p(face) under high noise conditions as compared to low noise conditions (gray lines). SEM indicated by error bars. See Figure S1 for behavioral group according to monkey.
(C) Estimates of p(face) over a single recording session. Estimates were derived using a reinforcement learning (delta rule) model (green line) and a Bayesian model (turquoise line) and are compared to the generative probability (black line). Ovals along the bottom of the figure indicate trials where the cue stimulus was a face.
Current Biology  , DOI: ( /j.cub )
Copyright © Terms and Conditions

3. Figure 2 Responses of Individual IT Neurons to Face and Fruit Stimuli
(A) Four examples of neurons from monkey 1 (left, middle left panel) and monkey 2 (middle right, right panel) to high (dashed lines) and low (solid lines) noise face (red) and fruit (blue) stimuli. In all four examples, low noise face stimuli evoked a strong and persistent response. Fruit stimuli and high noise evoked comparatively weaker responses.
(B) Responses of the same four neurons, with trials sorted according to face expectation. Solid lines show responses for trials where faces were expected (p(face) > 0.66); dashed lines show responses for trials where face and fruit were equally expected (0.66 > p(face) > 0.33); dotted lines show responses for trials where face were not expected (i.e., fruit stimuli were expected; p(face) < 0.33). All neuronal responses were smoothed with a 10-ms Gaussian kernel (see the Supplemental Experimental Procedures).
Current Biology  , DOI: ( /j.cub )
Copyright © Terms and Conditions

4. Figure 3
Population Responses in IT to Expected, Neutral, and Unexpected Faces
(A–D) Population responses in IT to expected, neutral, and unexpected faces (A and C) and fruit stimuli (B and D) for low and high noise trials. A robust and sustained increase in activity in response to face stimuli was observed, beginning ∼100 ms following cue onset (A). Black bar indicates cue presentation period. For (A) and (D), red/blue bars indicate time points where average firing rates for expected and unexpected faces/fruit were significantly different from one another (light bars, unexpected > neutral; dark bars, neutral > expected).
(E and F) Cue-aligned regression coefficients for the interaction between stimulus category and p(face) (β3) (E) and stimulus category (β1) for the regression analysis from Equation 1 (F). In both cases, three separate lines (largely overlapping) show β from separate regressions in which the previous 1 (dark red lines), 2 (medium red), or 3 (light red) trials were included as nuisance covariates. Neuronal responses were smoothed with a 10-ms Gaussian kernel (see the Supplemental Experimental Procedures). The red bars show the corresponding significant time points for each analysis. To correct for multiple comparisons, a cluster-correction was applied across time points (∗p < 0.05).
See also Figure S2.
Current Biology  , DOI: ( /j.cub )
Copyright © Terms and Conditions

5. Figure 4 Prediction Signals in IT Neurons
(A) Cue-aligned regression coefficients for prediction (β2), prediction error (β3), stimulus category (β1), and the interaction between stimulus category and previous choice (β4) all from Equation 4. Black bars indicate stimulus presentation period. Red bars show cluster-corrected significant time points.
(B) Split-half cross-validation of cue-aligned regression coefficients within each predictor variable. Each plot shows the Pearson’s correlation between regression coefficients for two independent splits of the data (x axis, y axis). Each panel shows correlations within a predictor variable (e.g., β2 and β2 in left panel) and for a given time point with every other time point. Points along the diagonal of each panel show correlations in selectivity among equivalent time points (e.g., 200 ms post-stimulus in both halves of the data), whereas off-diagonal points indicate correlations between differing time points (e.g., 100 ms post-stimulus in one half of the data, with 300 ms in the other). Black contour lines indicate time points where a significant correlation was observed (p < 0.05), adjusted for multiple comparisons using a cluster correction method. Dashed lines indicate the quadrants used for significance testing in the main text.
(C and D) Scatter plots between independently obtained parameter estimates from p(face)RL and Δp(face)RL for each neurons in the pre-stimulus (C) and post-stimulus (D) periods. Inset text shows the p value for the corrected non-parametric test (see the Supplemental Experimental Procedures). Red line shows the best-fitting linear trend for significant correlations.
See also Figure S3.
Current Biology  , DOI: ( /j.cub )
Copyright © Terms and Conditions

6. Figure 5
Correlations between Encoding of Faces, Face Predictions, and Face Prediction Errors in IT Neurons
(A) Cross-validation of cue-aligned regression coefficients. Each panel shows the Pearson’s correlation between two different predictor variables for two independent splits of the data (e.g., β1 and β2, left panel). Every time point is correlated with every other time point. Black contour lines indicate time points where a significant correlation was observed (p < 0.01), adjusted for multiple comparisons using a cluster correction method. Dashed lines indicate the quadrants used for significance testing in the main text.
(B and C) Correlation between parameter estimates for each neuron in the pre-stimulus (B) and post-stimulus (C) periods. Inset text shows the p value for the corrected non-parametric test (see the Supplemental Experimental Procedures).
See also Figures S4 and S5.
Current Biology  , DOI: ( /j.cub )
Copyright © Terms and Conditions

7. Figure 6 Encoding of Prediction Signals at the Time of Choice
(A) Regression coefficients (from Equation 5) for choice (β1), prediction (β2), and their interaction (β3) at the time the monkey made its choice (aligned on eye-on-target), for correct trials only (red lines) and incorrect trials only (blue lines). Average responses of IT neurons at time of choice shown in Figure S3. Cross-validation of choice-aligned regression coefficients shown in Figure S4.
(B and C) Decoding accuracy for the cue identity from cue (left panels) and response-aligned activity (right panels) for low-noise (B) and high-noise (C) trials. Decoding of cue identity was significantly above chance for the majority of the trial duration, in cases where a given cue was expected (i.e., p(face) > 0.66 and a face stimulus occurred, or where p(face) < 0.33 and a fruit cue occurred). This was not the case for neutral trials (i.e., p(face) > 0.33 and < 0.66), where cue identity could not be decoded until after cue onset. Red bars and black bars show time points where decoding was significantly above chance for the expected and neutral trials respectively, computed using a permutation testing method. Trial counts were equated for all analysis conditions.
See also Figure S6.
Current Biology  , DOI: ( /j.cub )
Copyright © Terms and Conditions